Governance framework for AI agents in enterprise networks. Exposes network operations tools via MCP protocol with policy-based access control, risk evaluation, and escalation workflows.
Both tools have critical definition gaps. READ_TOOLS and WRITE_TOOLS are generic collection names that violate verb_noun naming (pattern:tool). Descriptions are present but generic ('Collection of read-only network operations tools'). Input schemas exist with 4 parameters (device, confidence_score, context, caller_id) but lack critical details: no enum constraints on device names, no range/format specs for confidence_score (0.0-1.0 stated in description but not in schema), context is untyped object (should be structured), no output schema documented. No error handling guidance. Tools appear to be tool collections rather than individual tools, making composition unclear. Cannot verify actual tool registration in provided code snippet.
Collection of read-only network operations tools for querying device state and configuration
Collection of write operation tools for modifying device configuration and state
Tool names violate verb_noun pattern. 'READ_TOOLS' and 'WRITE_TOOLS' are collection names, not action verbs. LLMs cannot infer intent from these names. Should be specific actions like 'get_device_config', 'list_device_interfaces', 'update_device_config', 'send_device_command'.
Input schema lacks type definitions for 'context' parameter (declared as 'object' with no properties). Cannot validate or guide LLM on what context fields are valid. Should define context as a structured object with typed properties.
No output schema documented for either tool. LLMs cannot plan downstream calls or extract required fields. Must document what fields are returned, their types, and which IDs/references are included for tool chaining.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 30 | <=2025-11-25 | v2 |
confidence_score parameter has range constraint (0.0-1.0) stated only in description, not in schema. JSON Schema should include minimum/maximum. LLMs cannot read description-only constraints reliably.
No error handling guidance. Tools declare risk levels (READ_ONLY, WRITE) but provide no recovery instructions. If a device is unreachable or policy blocks execution, what should the LLM do next?